Source-linked AI summary
Generative artificial intelligence enhances creativity but reduces the diversity of novel content
Anil R. Doshi, Oliver P. Hauser
TL;DR
The paper examines whether access to GenAI ideas enhances or constrains creativity and studies this question in an online writing experiment. GenAI access causally increased stories’ novelty and usefulness, with larger gains among less creative writers, while raising concerns about reduced diversity.
Problem
The paper asks whether GenAI ideas enhance or constrain creative output, including the potential consequences of widespread reliance for creative diversity.
Source: Discussion · p. 12Method
An online experiment randomly varied whether writers received one or five GenAI story ideas, using writers without AI access as the baseline.
Source: Discussion · p. 10, Methods · p. 15Results
GenAI ideas increased story novelty by 6.7% and usefulness by 6.4%, with larger creativity gains among less creative writers.
Source: Results · p. 4, Discussion · p. 11Takeaways & Limitations
GenAI can improve individual creative outputs, but broader adoption may make produced stories less unique and more similar to one another.
Source: Discussion · pp. 11–12Because writers could not customize GenAI prompts or interact repeatedly with the system, the estimates may be lower bounds of its potential impact.
Source: Discussion · p. 11
Abstract
from arXiv · showhide
Creativity is core to being human. Generative artificial intelligence (GenAI) holds promise for humans to be more creative by offering new ideas, or less creative by anchoring on GenAI ideas. We study the causal impact of GenAI on the production of a creative output in an online experimental study where some writers are could obtain ideas for a story from a GenAI platform. Access to GenAI ideas causes an increase in the writer's creativity with stories being evaluated as better written and more enjoyable, especially among less creative writers. However, GenAI-enabled stories are more similar to each other than stories by humans alone. Our results have implications for researchers, policy-makers and practitioners interested in bolstering creativity, but point to potential downstream consequences from over-reliance.
Results
Access to GenAI ideas improved third-party evaluations of story creativity and emotional qualities, especially for less inherently creative writers, but reduced ownership and increased similarity among stories. Evaluators generally accepted AI-assisted creativity while expecting acknowledgment and compensation, whereas writers’ self-assessments showed no significant condition differences.
Baseline versus both GenAI combined
- GenAI access increased stories’ evaluated novelty by 6.7% and usefulness by 6.4% versus Human only.These effects were supported across the constituent novelty and usefulness components.Source: Results · p. 4, Figure 2 · p. 4
Variation by GenAI conditions
- Access to five GenAI ideas produced larger gains than one idea, increasing novelty by 8.1% versus 5.4% and usefulness by 9.0% versus 3.7%.Participants requested GenAI ideas frequently: 88.4% used GenAI at least once, and five-idea participants requested 2.55 ideas on average.Source: Results · p. 5
Exploratory analyses: emotional characteristics
- Third-party evaluators found GenAI-assisted stories more enjoyable, more likely to contain plot twists, better written, less boring, and more expectation-changing, but not funnier.Writers’ self-assessments showed no significant differences across emotional characteristics or creativity outcomes.Source: Results · pp. 5–6
Heterogeneity by inherent creativity
- GenAI access had little effect on high-DAT writers but substantially improved low-DAT writers’ outcomes, with five ideas raising low-DAT novelty by 10.7% and usefulness by 11.5%.For low-DAT writers, five ideas increased evaluations of writing quality by up to 26.6% and enjoyment by up to 22.6%, effectively equalizing creativity scores.Source: Results · pp. 6–7, Figure 3 · p. 7
Ownership and profit-sharing
- Evaluators attributed 25.4% less ownership to writers with one GenAI idea and 31.0% less ownership to writers with five ideas than to Human-only writers.Each GenAI condition was identifiable to evaluators, while five ideas produced only a marginally significant 2.3% profit penalty relative to one idea; ownership strongly predicted profit allocation.Source: Results · pp. 7–8, Figure 4 · p. 8
GenAI and ethics in the creative process
- Evaluators generally viewed GenAI story use as ethical and creative, accepted AI for initial ideas, but opposed undisclosed use and supported compensating source content creators.They also tended to support disclosure of AI-generated content alongside the final story.Source: Results · p. 9
Similarity of stories
- Stories with one or five GenAI ideas were 5.2% and 5.0% more similar to GenAI ideas than Human-only stories, respectively, indicating anchoring.A 4.29-point increase for one GenAI idea represented 24.3% of the Human-only similarity range.Source: Results · p. 9, Figure 5 · p. 10
- GenAI-assisted stories were also more similar to the average of other stories in their condition, with coefficients of 0.871 for one idea and 0.718 for five ideas.This indicates reduced diversity among stories produced with GenAI access.Source: Results · pp. 9–10, Figure 5 · p. 10
Discussion
The study finds that access to GenAI causally enhances average creativity in short-story writing, while warning that broader adoption may reduce the diversity of creative outputs. These effects are strongest for less creative writers, but the study’s controlled design and participant sample limit how broadly the findings can be generalized.
Main findings
- 10% to 11% increases in creativity and 22% to 26% increases in enjoyability and writing quality show that GenAI improves stories, especially for less creative writers.Access to up to five alternative GenAI ideas increased average novelty and usefulness relative to writing without GenAI.Source: Discussion · pp. 10–11
Heterogeneous effects
- GenAI equalizes story evaluations by helping less creative writers, but the study finds no evidence that it raises the creative ceiling of especially creative humans.Future research should test deeper interaction with specialized tools to determine whether more creative participants can benefit further.Source: Discussion · p. 11
Study design and limitations
- The tightly controlled design supports causal interpretation but likely understates GenAI’s potential because writers could not customize prompts or interact repeatedly with the system.Nearly 9 out of 10 participants offered GenAI ideas requested at least one, while estimates used a conservative intention-to-treat analysis.Source: Discussion · p. 11
Scope and generalizability
- The study examines relatively typical participants rather than professional writers or unusually creative individuals, limiting population representativeness while suggesting the largest effects may occur among less creative people.Professional and unusually creative populations may experience different efficiency, speed, or creative effects.Source: Discussion · pp. 11–12
Aggregate diversity
- Despite improving average creativity, widespread use of GenAI-inspired stories could make outputs less unique and more similar to one another, potentially reducing diversity over time.The concern may intensify as models train on existing and AI-generated text, with initial evidence that models trained on GenAI content become unstable.Source: Discussion · p. 12
Future technology
- Because GenAI is rapidly evolving, GPT-4 may soon become obsolete, creating opportunities to test how newer systems affect creativity, art, literature, and innovation.The authors view current enhancement effects as a starting point rather than a limit on future creative possibilities.Source: Discussion · p. 12
Methods
The study randomized writers to human-only or GenAI-assisted story writing, then evaluated 293 resulting stories through independent reader assessments and preregistered outcome measures. Analyses used OLS with robust or evaluator-clustered standard errors, comparing GenAI conditions against human-only writing.
Writer study
- After exclusions, the writer study contained 293 writers and stories recruited from 500 initial Prolific participants.Participants were UK-based Prolific users with approval ratings of at least 95%; exclusions included missing consent, dropout, and admitted GenAI use in the Human only condition.Source: Methods · p. 12
- Writers were randomized to Human only, Human with 1 GenAI idea, or Human with 5 GenAI ideas before producing exactly eight-sentence stories on one of three adventure topics.GenAI ideas were optional, generated by GPT-4 from topic-specific prompts, and could not be copied and pasted.Source: Methods · p. 13
Evaluator study
- Independent Prolific evaluators assessed the 293 stories for stylistic characteristics, novelty, usefulness, perceived AI authorship, ownership, profits, and ethical views.The evaluator study produced 3,519 evaluations by 600 evaluators, with each evaluator initially shown six stories.Source: Methods · pp. 13–14
Outcome variables
- Creativity was preregistered as separate novelty and usefulness indices, each averaging three 9-point components, alongside exploratory stylistic outcomes.Novelty comprised novel, original, and rare; usefulness comprised appropriate, feasible, and publishable.Source: Methods · p. 14
Similarity scores
- Similarity was measured from 0 to 100 using cosine similarity between story embeddings and GenAI ideas or the average embedding of other stories in the same condition.Embeddings came from OpenAI’s embeddings API, with randomly assigned ideas used when writers lacked an idea and the first idea used for multiple-idea writers.Source: Methods · p. 15
Statistical analysis
- OLS regressions used robust standard errors for writer outcomes and evaluator-clustered robust standard errors, with Human Only as the baseline for exogenously assigned GenAI conditions.The study was preregistered at AsPredicted.org (ID 136723) and approved by UCL and the University of Exeter ethics boards.Source: Methods · p. 15
Supplemental Information for: Generative artificial intelligence enhances creativity
The supplemental information documents the study instruments, supporting analyses, preregistered materials, and study interfaces. Its substantive reported result is that stories from Human-only participants resembled stories from GenAI-condition participants who did not use AI more than stories from participants who used assigned GenAI ideas.
Section 1. Summary of questions for Writer and Evaluator studies
- The supplement includes questions about story novelty, prospects for expansion into a book, ownership, and compensation when AI-generated output draws on others’ content.Participants indicate agreement or responses on a 1 (not at all) to 9 (extremely) scale.Source: Section 1. Summary of questions for… · pp. 19–20
Section 2. Supporting Tables and Figures
- Supporting tables and figures cover writer and evaluator means, creativity and emotional outcomes, robustness checks, interactions with DAT scores, AI assistance, ownership, profits, survey responses, and story similarity.The supporting materials include 17 supplementary tables and figures on these outcomes and analyses.Source: Table 1 · p. 22, Table 2 · p. 23, Table 5 · p. 25
- Stories from Human-only participants and GenAI-condition participants who did not use AI had more similar similarity distributions and were less similar to assigned GenAI ideas than stories from participants who used GenAI ideas.The analysis used simulated AI ideas for the first two groups and compared their story–idea cosine similarity with participants who accessed an actual GenAI idea.Source: Section 2. Supporting Tables and Figures · p. 35, Figure 4 · p. 36, Table 18 · p. 36
Section 5. Writer study screenshots
- Writer-study screenshots document conditions providing one or five AI ideas before or after story generation, alongside condition-specific pages and prompts.The screenshots also distinguish Human-only participants and participants shown follow-up questions based on prior responses.Source: Section 5. Writer study screenshots · pp. 44, 45, 46, 48, 56
Section 6. Evaluator study screenshot
- Evaluator-study screenshots provide example stories from Human-only, non-accessing, one-idea, and five-idea conditions, including cases where writers reported outside AI use.The examples distinguish requested access to one or five GenAI ideas from stories produced without accessing the offered idea.Source: Section 6. Evaluator study screenshot · pp. 70–74